1. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Multimodal fitting of atypical size distributions from AERONET Michael Taylor Stelios Kazadzis Evangelos Gerasopoulos URL: http://apcg.meteo.noa.gr eMail: mtaylor@noa.gr OVERVIEW 1.Typical & atypical distributions1.Typical & atypical distributions 2.Two new fitting methods2.Two new fitting methods 3.An interesting new case3.An interesting new case 4.Potential impact & a wish list 4.Potential impact & a wish list OVERVIEW 1.Typical & atypical distributions1.Typical & atypical distributions 2.Two new fitting methods2.Two new fitting methods 3.An interesting new case3.An interesting new case 4.Potential impact & a wish list 4.Potential impact & a wish list

2. 1. Typical & atypical distributions Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session

3. Fresno (8 th Feb 2006) 0.439 ≤ r s ≤ 0.992μm AERONET mode separation range rfrf rcrc σfσf σcσc Fine Mode: V f Coarse Mode: V c 1a) A typical (bi-modal) size distribution1a) A typical (bi-modal) size distribution ε=35% ε=10% ε=100%

4. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session 1a) 2 days later…1a) 2 days later… r=0.19μmr=0.44μm r=0.26μm 1.mis-identification of “fine” modes 2.creation of a “ghost” (non-physical) fine mode 3.drop in goodness of fit: R 2 =0.997  R 2 =0.892 1.mis-identification of “fine” modes 2.creation of a “ghost” (non-physical) fine mode 3.drop in goodness of fit: R 2 =0.997  R 2 =0.892 Fresno (10 th Feb 2006)

5. 1a) another typical (bi-modal) size distribution1a) another typical (bi-modal) size distribution Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Lanai (11 th Jan 2002)

6. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session 1a) 10 days later…1a) 10 days later… r=1.35μm r=4.1μm r=1.92μm 1.mis-identification of “coarse” mode(s) 2.creation of a “ghost” (non-physical) coarse mode 3.drop in goodness of fit: R 2 =0.934  R 2 =0.885 1.mis-identification of “coarse” mode(s) 2.creation of a “ghost” (non-physical) coarse mode 3.drop in goodness of fit: R 2 =0.934  R 2 =0.885 Lanai (21 st Jan 2002)

7. 1b) IDEA: an initial taxonomy of atypical distributions1b) IDEA: an initial taxonomy of atypical distributions Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Lanai (21 st Jan 2002) Double-coarse peak Lanai (21 st Jan 2002) Double-coarse peak Fresno (10 th Feb 2006) Double-fine peak Fresno (10 th Feb 2006) Double-fine peak Washington-GSFC (23 th Jun 1993) Triple peak (Pinatubo ash effect) Washington-GSFC (23 th Jun 1993) Triple peak (Pinatubo ash effect) Solar Village (29 th Mar 2011) Quenched fine mode Solar Village (29 th Mar 2011) Quenched fine mode Beijing (18 th Feb 2011) Skewed fine & coarse peaks Beijing (18 th Feb 2011) Skewed fine & coarse peaks Beijing (23 th Feb 2011) Elevated mid-point Beijing (23 th Feb 2011) Elevated mid-point Q. Anyone interested in collaborating to extend our database of atypical events ?

8. 1c) IDEA: can we use R 2 to detect atypical events?1c) IDEA: can we use R 2 to detect atypical events? Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session AERONET SiteDateObservationR2R2 GSFC-Washington23-Jun-93Triple Peak0.309 Lanai20-Jan-02Double Coarse Peak0.833 Fresno10-Feb-06Double Fine Peak0.892 Beijing18-Feb-11Skewed Fine & Coarse Peaks0.894 Beijing23-Feb-11Elevated Mid-Point0.924 Solar Village29-Mar-11Quenched Fine Peak0.962 Frenso08-Feb-06Bi-modal0.997 Typical case Strongly atypical Moderately atypical Weakly atypical

9. 2. Two new fitting methods Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Taylor, Kazadzis, Gerasopoulous (2014): AMT 7, 839-858

10. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session 2a) Optimized Equivalent Volume (OEV) method2a) Optimized Equivalent Volume (OEV) method Lanai (21 st Jan 2002) Vary r s

11. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session 2a) OEV method: varying r s Max(R 2 )  criterion for identifying optimal r s

12. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session 2a) OEV method: comparison with AERONET bi-modal2a) OEV method: comparison with AERONET bi-modal TINY quantitative but not qualitative improvement Lanai (21 st Jan 2002)

13. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session 2b) Gaussian Mixture Model (GMM) method varying n-modes2b) Gaussian Mixture Model (GMM) method varying n-modes Lanai (21 st Jan 2002)

14. n ModesR 2 (n)R 2 (n+1)F(ρ 1 )F(ρ 2 )CI 1 (l)CI 1 (u)CI 2 (l)CI 2 (u)t-Welch 10.777 2 0.8191.381.501.341.541.461.543.87 30.8190.9981.503.801.463.843.763.8476.26 40.9980.9933.803.173.763.213.133.2120.80 Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session 2b) GMM method: the stopping condition2b) GMM method: the stopping condition Fisher (1921): Metron 1: 3–32 Welch (1947): Biometrika 34 (1–2): 28–35 (maximum RE=0.060%) (95% confidence level) CASE 2: In the event of an overlap, there is a statistical difference when t>1.96 Harel (2009): App Stat 36(10): 1109-1118

15. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session 2b) GMM method: residual analysis2b) GMM method: residual analysis LARGE reduction in residuals when n = 3 and a small increase when n = 4 Lanai (21 st Jan 2002)

16. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session 2b) GMM method: comparison with AERONET bi-modal2b) GMM method: comparison with AERONET bi-modal LARGE quantitative and qualitative improvement Lanai (21 st Jan 2002)

17. 3. An interesting new case Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session

18. 3a) Fresno: 8 th –14 th Feb 2006: MODIS 2km + aerosol props.3a) Fresno: 8 th –14 th Feb 2006: MODIS 2km + aerosol props. 11 th Feb 2006 13 th Feb 2006 14 th Feb 2006 9 th Feb 2006 10 th Feb 2006 8 th Feb 2006 AERONET AOD Product Level 2 (Version 2) GOCART V4 %AOD per aerosol type Almost NO change in composition but a LARGE change in H2O (x2) Almost NO change in composition but a LARGE change in H2O (x2)

19. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session 8 th Feb 9 th Feb 10 th Feb 11 th Feb 13 th Feb 14 th Feb 3a) Fresno 8 th -14 th Feb 2006  Fog-induced modification3a) Fresno 8 th -14 th Feb 2006  Fog-induced modification

20. 4. Potential impacts Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session

21. 4b) Potential impact: on AERONET & knock-on effects?4b) Potential impact: on AERONET & knock-on effects? Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Regional models & GCMs AERONET inversions AERONET inversions SATELLITE retrievals SATELLITE retrievals LIDAR profiles LIDAR profiles IN SITU chemistry IN SITU chemistry AERONET is a reference for: 1)validating satellite retrievals 2)assigning aerosol types 3)providing 1 st guesses in “spin-ups” AERONET is a reference for: 1)validating satellite retrievals 2)assigning aerosol types 3)providing 1 st guesses in “spin-ups”

22. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session 4) A wish list4) A wish list 1)More continuity in the AERONET inversion data record  to enable studies of the temporal evolution of atypical aerosol events 2)Establishment of a taxonomy database  to help detect, assess and monitor atypical events 3)Incorporation of our algorithm into operational algorithms & AERONET (inversion) data products 4)Your suggestions 1)More continuity in the AERONET inversion data record  to enable studies of the temporal evolution of atypical aerosol events 2)Establishment of a taxonomy database  to help detect, assess and monitor atypical events 3)Incorporation of our algorithm into operational algorithms & AERONET (inversion) data products 4)Your suggestions Many thanks to all our colleagues Michael Taylor, IERSD-NOA mtaylor@noa.gr

23. EXTRA SLIDES Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session

24. Errors on the AERONET retrievalErrors on the AERONET retrieval Lanai (21 st Jan 2002) ε=100% ε=10% ε=35% ε=100% ε = reported AERONET error: ε = 10% at fine & coarse peaks ε = 35% at local minimum ε = 100% at edges: r 7μm Dubovik et al (2002) ε = reported AERONET error: ε = 10% at fine & coarse peaks ε = 35% at local minimum ε = 100% at edges: r 7μm Dubovik et al (2002) Lanai (21 st Jan 2002) N=2200 (x10 AERONET) gave best results for: 1)calculation of max(R 2 ) in the OEV method 2)stabilization of the SE in the GMM method N=2200 (x10 AERONET) gave best results for: 1)calculation of max(R 2 ) in the OEV method 2)stabilization of the SE in the GMM method

25. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Using GOCART to isolate eventsUsing GOCART to isolate events Source: IPCC/AR5 (2013) GOCART 2000-2006 global mean AOD per type

26. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Aerosol typeAERONET sitePeak dateSUDUSSOC+BC DustBanizoumbou16/03/20051.02%97.91%0.03%1.04% Biomass BurningMongu14/08/20035.61%0.22%0.05%94.12% Urban SO 2 GSFC-Washington17/08/200587.53%1.38%0.05%11.04% Marine (sea salt)Lanai21/02/200228.92%3.32%60.14%7.61% Aerosol typep [r f ]p [r c ]p [σ f ]p [σ c ]p [Vf]p [V c ] Dust0.014 **0.0390.4560.0830.013 ** Biomass Burning0.1600.9300.0920.1780.6540.678 Urban SO 2 0.5720.9820.2370.1200.1390.152 Marine (sea salt)0.017 **0.0350.0480.1320.012 **0.008 ** **  statistically-significant for dust & marine-dominated AVSD (2-tailed paired t-test at the 95% level of confidence: p<0.025) **  statistically-significant for dust & marine-dominated AVSD (2-tailed paired t-test at the 95% level of confidence: p<0.025) OEV method: “pure” aerosol casesOEV method: “pure” aerosol cases

27. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session GMM V OEV V AERONET: “pure” aerosol casesGMM V OEV V AERONET: “pure” aerosol cases SEA SALT (60%): Lanai (21 st Jan 2002) = “Double coarse peak” DUST (98%): Banizoumbou (16 th March 2005) = “Quenched fine mode” URBAN SO 2 (88%) : GFSC-Washington (17 th August 2005) = “Typical” SMOKE (94%): Mongu (14 th August 2003) = “Typical”

28. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Derivatives are sensitive to N Derivatives are sensitive to N GMM method: some more mathsGMM method: some more maths

29. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Potential impact: on AERONET optical propertiesPotential impact: on AERONET optical properties Lanai (Jan 2002) V f = 0.12 x AOD(1020)  AOD(1020) ≈ 8.33 V f V c = 0.93 x AOD(1020)  AOD(1020) ≈ 1.08 V c r s (AERONET) = 0.439µm but r s (OEV) = 0.587µm  RE(AERONET-OEV) ≈ -28% for V f and ≈ +7% for V c  AOD(1020) for the fine mode ≈ 28% higher with OEV than AERONET  AOD(1020) for the coarse mode ≈ 7% lower with OEV than AERONET

30. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Global PM10 concentrationsGlobal PM10 concentrations Source: IPCC/AR5 (2013)

31. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Fine & coarse mode chemistryFine & coarse mode chemistry After: Gerasopoulos et al (2007)

32. Aerosol -radiation-cloud uncertainty is confidently LARGEAerosol -radiation-cloud uncertainty is confidently LARGE Source: IPCC/AR5 (2013) Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Accurate V Precise

33. Michael Taylor, COMECAP 29 th May, 2014: Remote Sensing Session Tropospheric aerosol lifetimesTropospheric aerosol lifetimes Source: IPCC/AR5 (2013)